Why Your AI Agent Has Terrible Memory (And the Simple Fix Most People Miss)
"It keeps forgetting everything."
This is the most common complaint from people who've tried OpenClaw and almost given up. They've done the install, connected Telegram, started having conversations — and two days later, the agent acts like it's never met them. No memory of their projects. No retention of preferences. A clean slate every single session.
The good news is that this is not a model limitation. It's not an inherent flaw in how AI works. It's a configuration problem, and it has a direct fix that takes less than 30 minutes to implement.
Here's exactly what's going wrong and how to resolve it permanently.
The Root Cause: A File That Doesn't Exist
When OpenClaw installs, it creates your workspace folder with several important configuration files. What it doesn't create by default is MEMORY.md — the file that's supposed to serve as your agent's long-term memory.
OpenClaw only auto-loads a specific set of files into every session: SOUL.md, AGENTS.md, USER.md, TOOLS.md, IDENTITY.md, HEARTBEAT.md, BOOTSTRAP.md, and MEMORY.md. These eight files are loaded automatically at the start of every conversation. Everything in them, your agent knows.
The problem is that MEMORY.md is on this auto-load list, but it doesn't exist until you create it. So your agent has nowhere to write. Nothing persists. Every session starts fresh.
The fix is immediate: tell your agent to create the file. Open your OpenClaw chat and say: "Create a MEMORY.md file in my workspace and start using it to log important information from our sessions." That's it. The file exists now, and so does the mechanism for persistence.
How Memory Actually Works in OpenClaw
Understanding the system makes it easier to configure correctly. There are three layers.
Layer 1: Session Files (Always Active)
Every session, the eight auto-load files are pulled into context before you say anything. Your agent knows your name, your projects, your behavior preferences, and your past decisions — all from these files. Managing them well is the foundation of having an agent that feels like it knows you.
Layer 2: Long-Term Memory (MEMORY.md)
This is the high-level memory file. It should contain durable information: things you've explicitly told your agent to remember, decisions you've made, preferences that should persist forever, key people and projects in your life. Think of it as the agent's personal notes about you — the stuff that should never disappear.
The agent writes to MEMORY.md when instructed to, either explicitly by you or by rules you've put in AGENTS.md or SOUL.md. A good rule to add to SOUL.md: "When the user shares important decisions, preferences, or context, write them to MEMORY.md immediately."
Layer 3: Daily Memory Logs (/memory folder)
More granular than MEMORY.md, the /memory folder contains one file per day — named YYYY-MM-DD.md — with running context and observations from that day's sessions. Today's and yesterday's files are automatically loaded into every session, giving your agent short-term context about recent work.
After 30 days of consistent logging, this folder becomes a searchable archive of everything you've worked on. After 90 days, your agent understands your patterns well enough to anticipate your needs.
The Context Window Compaction Problem
Even with both files in place, there's a second failure mode: context compaction.
As your session grows longer, OpenClaw eventually approaches the context window limit and performs a compaction — essentially summarizing everything to free up space. Information gets lost in this process. Nuance gets flattened. Specific details disappear.
Two settings prevent this:
set compaction_memory_flush_enable to true
set memory_search.experimental.session_memory to trueWhat these do: force your agent to write everything important to memory before the compaction happens. Nothing gets lost in the summarization because it was already saved.
Run both commands in your OpenClaw terminal. Some newer versions have these enabled by default, but it's worth confirming.
The Heartbeat Auto-Save Solution
Even with the files created and the settings enabled, memory can still slip through the cracks — especially if you end sessions abruptly or switch context quickly.
The most reliable fix is adding an auto-save instruction to your heartbeat file. The heartbeat runs every 30 minutes regardless of activity, which makes it a perfect safety net for memory.
Add this to your HEARTBEAT.md:
MEMORY MAINTENANCE
Check if today's memory file exists at /workspace/memory/YYYY-MM-DD.md.
If it doesn't exist, create it.
Summarize key discussions, decisions, and important context from all active sessions.
Write any important user preferences or decisions to MEMORY.md.Now every 30 minutes, your agent is actively logging what it knows. Memory becomes reliable infrastructure rather than a feature that sometimes works.
Diagnosing Memory Problems
If you're not sure whether memory is loading correctly, run /context list in your OpenClaw session. This shows exactly which files are currently loaded into context. If MEMORY.md shows as missing or isn't listed, it's not in context — and that's why your agent appears to forget things.
Other common issues:
Symlinks breaking memory. If your workspace files are symlinks pointing outside the workspace root, OpenClaw's security check blocks them from loading. Replace symlinks with real file copies.
Agent not writing to memory automatically. You may need to explicitly instruct it. Add rules to AGENTS.md or tell it directly: "When I tell you something important, save it to MEMORY.md right now."
Memory file getting too large. After several months, MEMORY.md can become very long. Review it periodically and consolidate or archive older entries. A bloated memory file is better than no memory file, but a clean one loads faster.
What Good Memory Looks Like After 30 Days
A properly configured OpenClaw agent should accumulate, over time: your communication preferences and tone, recurring projects with full context, the people you work with and your relationship to them, decisions you've made and the reasoning behind them, things you've explicitly asked it to remember, and things you've asked it to stop doing.
Thirty days in, your agent feels like a colleague who's been with you for months. Sixty days in, it starts to feel like the best assistant you've ever had. The compounding is real — but only if the memory infrastructure is working.
The Quick Checklist
- Create
MEMORY.mdin your workspace if it doesn't exist - Create the
/memoryfolder for daily logs - Add the memory save rule to
SOUL.mdorAGENTS.md - Run the two compaction flush settings
- Add memory maintenance to your heartbeat file
- Run
/context listto verify files are loading - Review and clean
MEMORY.mdmonthly
Five of these steps take under two minutes each. Memory stops being a problem before the end of the day.
Based on a conversation with Moritz Kram on the Startup Ideas podcast.
